{"id":"23a1677c-3418-4178-888d-86560a0e92ed","arxiv_id":"2507.18275","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Dis-GEN learns to represent and generate disordered inorganic crystal structures with partial atomic occupancy and vacancies while preserving crystallographic symmetry.","lead":"Dis-GEN is an AI model that generates crystal structures where different atoms share the same sites, a common feature of real battery and semiconductor materials. It uses a new way of describing these messy crystals, so symmetry is preserved while allowing partial occupancy and vacancies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The validity filter used to certify generated structures is undefined for partial occupancies, so Table 3 does not support the central generation claim.","rationale":"The reader's conditional verdict is reasonable. Reconstruction errors are credible, but the generation evaluation has a concrete logical gap. The no-overlap criterion is adopted unchanged from a model that outputs fully occupied ordered structures; applying it to partial-occupancy CIFs requires a decision (average positions, dominant species, supercell enumeration) that is not documented. Any of those choices changes pass rates dramatically. This is not merely a missing baseline or missing code; it is an internal inconsistency between the data representation and the evaluation metric. The paper's own admission of chemically unreasonable generations and unstable Zn-V-O candidates reinforces this: the filters that supposedly ensure chemical reasonableness are not doing that work. The fix is straightforward: define validity on a representative supercell or on the average structure with an explicit cutoff. Until then, Table 3 does not support the strongest claim. Since the representation and reconstruction results could still hold, a conditional verdict remains appropriate, and no verdict change is needed.","tokens_in":16770,"tokens_out":4384,"duration_ms":49582,"concrete_test":"Obtain Dis-GEN's released generation code or generated CIFs and run pymatgen's nearest-neighbor distance on the output CIFs exactly as generated, preserving partial occupancies as in Fig. 1. If any Wyckoff site has two or more species at identical coordinates, the structure fails the stated 0.5 Å cutoff; if the implementation collapses or omits species before measuring, then the validity metric is not applied to the disordered degrees of freedom. Recompute Table 3 under both readings and report the fraction of KDE samples that pass the criterion on the actual generated representation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing weakness is the structural-validity filter in Section 4.3. It adopts Ref. 3's criterion that 'no two atoms be closer than 0.5 Å' and reports 96.44% validity for KDE generation. However, a disordered Wyckoff site by construction places several atomic species at the same fractional coordinates (Fig. 1: La and Sr both at (0,0,0.25)). Under the literal criterion, such a site has zero interatomic distance, so a genuinely disordered structure could not pass. Because the paper does not specify how partial occupancies are treated before applying the cutoff (collapse to a majority species, random supercell expansion, or deletion), the reported validity numbers do not test the generated partial-occupancy representation at all. This matters because SMA only checks Wyckoff letter/multiplicity consistency, and the remaining filter, charge neutrality, passes only about 55% of KDE samples; the paper itself flags generated compositions as 'chemically unreasonable' (Section 4.4) and shows Zn-V-O phases are thermodynamically unstable (Section 4.5). With the validity criterion unspecified, Table 3's headline '96.44% validity' cannot be interpreted, and the claim that Dis-GEN generates physically plausible disordered crystals rests on an unevaluated filter.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces Dis-GEN, a variational autoencoder for crystalline inorganic materials that include compositional disorder, i.e., sites with partial occupancies. The representation encodes each Wyckoff site by one-hot occupancy, multiplicity, Wyckoff letter, fractional coordinates, and a disorder indicator, together with lattice parameters and a one-hot space group. The model is trained on 138,692 structures filtered from ICSD, roughly half of which contain partial occupancies. The authors report low reconstruction errors on a held-out test set (lattice MAE ≈ 0.06–0.10 Å, space-group accuracy 99.4%, Wyckoff-letter/multiplicity accuracy 99.5–99.8%, partial-occupancy MAE 0.06). They then sample the latent space using a Gaussian, GMM, or KDE estimator and evaluate the sampled structures with symmetry-matching accuracy, a structural-validity cutoff, and a charge-neutrality filter. Conditional generation is demonstrated for a Zn-V-O composition-targeted search, but the generated phases are found to be thermodynamically unstable.","tokens_in":17050,"tokens_out":10374,"duration_ms":97110,"significance":"The core reconstruction experiments are clean and non-circular: a held-out ICSD split is encoded and decoded, and the errors on lattice parameters, space group, Wyckoff letters, multiplicities, and partial occupancies are small. These results support the less ambitious claim that a Wyckoff-site-level representation of disordered crystals with partial occupancies can be learned with a VAE. The use of experimental ICSD data is a genuine advantage over models trained only on DFT-relaxed ordered structures, and the geometric insight of grouping symmetry-equivalent sites is appropriate. If the evaluation issues were resolved, the representation could be a useful building block for generative modeling of disordered materials. However, the manuscript's central novelty claim—that Dis-GEN 'effectively generates' disordered inorganic crystals—is not yet supported: the generation-success metric is ambiguous, the validity metric is undefined for partial occupancies, charge neutrality passes only about 55%, and the one quantitative downstream test (Zn-V-O) yields unstable phases. No code or data are provided at submission, so the quantitative claims are not independently checkable.","major_comments":[{"comment":"The structural-validity metric is undefined for the partial-occupancy representation, so the reported validity percentages cannot be interpreted. The manuscript adopts Ref. 3's criterion that no two atoms be closer than 0.5 Å, but in Dis-GEN's representation a disordered Wyckoff site has several atomic species at identical fractional coordinates (e.g., La and Sr at (0,0,0.25) in Fig. 1). Under the literal rule, such a site has zero interatomic distance and would always fail the check. The paper does not state how partial occupancies are treated before the cutoff is applied—whether all species are retained, occupancies are collapsed to a majority species, or a random configurational supercell is constructed. Without this specification, the 96.44% KDE validity and 97.53% test-set validity provide no evidence about the physical plausibility of the generated structures. Since this is the main 'validity' column in Table 3, the generation claim rests on an unevaluated filter.","section":"4.3, Table 3"},{"comment":"The main generation metric is ambiguous. Section 4.2 defines the generation error as the percentage of sampled structures discarded during the reconstruction process, which gives 99.94% for KDE sampling. The Table 3 caption, however, says the generation error corresponds to 'the accuracy, defined as the fraction of structures that passed each respective filter,' which would mean 99.94% of KDE samples passed. These definitions point in opposite directions, and the manuscript never reports the joint success rate of reconstruction × SMA × validity × charge neutrality. If the discard interpretation is intended, only about 0.06% of KDE samples survive reconstruction, and combining with the 55.28% charge-neutrality rate gives an end-to-end success below 0.05%, so 'effective generation' is not demonstrated. If the pass interpretation is intended, the term 'generation error' and the Section 4.2 definition are misstated. The paper must correct this inconsistency and report the joint success rate explicitly.","section":"4.2–4.3, Table 3"},{"comment":"Charge neutrality and chemical reasonableness are not established. The charge-neutrality filter passes only 50–56% of structures for all estimators and the test set, meaning nearly half of the surviving generated structures are not charge-balanced. The paper's explanation that this is due to multi-valence species partially explains the difficulty, but it does not change the fact that the generation pipeline fails a basic chemical-plausibility screen. This is compounded by the admission in Section 4.4 that some generated compositions (e.g., Na7.92Mg8H3.76F10.88) 'appear chemically unreasonable.' These facts, together with the weak SMA check, mean the evidence supports reconstruction accuracy but not physically plausible generation of disordered crystals.","section":"4.3, Table 3; 4.4"},{"comment":"The conditional-generation case study undercuts the discovery claim. The Zn-V-O conditioned sampling produces structures that the paper itself reports are thermodynamically unstable: 'Both the ordered and disordered phases generated by Dis-GEN are found to be thermodynamically unstable' (Section 4.5). The attribution of this failure to the absence of property conditioning does not repair the issue; it confirms that the current model lacks the physical constraints needed for the proposed applications. The final claims in Section 5 that Dis-GEN 'enables crystal structure prediction for disordered crystals' and initiates 'systematic exploration of disordered inorganic crystals' are therefore too strong. The conclusions should be limited to the demonstrated representational learning and symmetry-consistent sampling, with conditional generation presented as a proof of concept that needs additional physical/chemical validation.","section":"4.5, Fig. 6, Section 5"}],"minor_comments":[{"comment":"There is a typo in 'determined by the sapce group'—it should read 'space group.'","section":"4.1"},{"comment":"The phrase 'rare atoms with a periodic number higher than 100' should be 'atomic number' rather than 'periodic number.'","section":"3.1"},{"comment":"There are several typos: 'Kullback-Leiber' should be 'Kullback-Leibler,' 'utalized' should be 'utilized,' and 'acknowlegde' should be 'acknowledge.'","section":"Appendix A, Appendix D, Acknowledgements"},{"comment":"The caption states that the decoder 'yields two two outputs'; this should read 'yields two outputs.'","section":"Figure 7 caption"},{"comment":"The SMA metric is an internal consistency check between predicted Wyckoff letter, multiplicity, and space group; it should not be described as evidence of symmetry preservation without also comparing the predicted symmetry attributes to the input structure's known values.","section":"4.1"},{"comment":"The statement that data and code 'will be public upon release' gives no repository or timeline, which makes it impossible to reproduce the reported numbers independently.","section":"Data and code accessibility"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a solid reconstruction core but the generation claims outrun the evidence. I would encourage the editor to require a clarified and unambiguous Table 3, a precise description of the validity protocol for partial occupancies, and a toned-down abstract and conclusion. The 'world's first' claim is not supported by a systematic comparison with existing methods that address disorder, and the lack of code or data at submission further weakens the reproducibility case. These issues are fixable within the scope of the paper, hence major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The representation is genuinely new: partial occupancies encoded per Wyckoff site, with a disorder indicator, one-hot Wyckoff letter/multiplicity, and space group/lattice vector. Reconstruction on held-out ICSD is clean: lattice MAE 0.06–0.10 Å, space group 99.4%, Wyckoff letter/multiplicity ~99.5%, partial occupancy MAE 0.06. That is real evidence the representation is learnable, and it is the paper's actual contribution.\n\nThe generation claim, though, is not established. The 96.44% validity number in Table 3 is uninterpretable as written. The criterion is 'no two atoms closer than 0.5 Å'; a disordered Wyckoff site has several species at identical fractional coordinates (Fig. 1: La and Sr both at (0,0,0.25)), so the literal criterion would give zero distance and fail every disordered site. The paper does not specify whether partial occupancies are collapsed, deleted, or expanded into a supercell before the cutoff. Without that, the validity metric tests something undefined. SMA checks Wyckoff letter/multiplicity against space group, which is an internal consistency check, not physical plausibility. Charge neutrality passes only ~55% of KDE samples, the paper itself calls some generated compositions chemically unreasonable, and the Zn-V-O case study shows generated phases are thermodynamically unstable. Those are the paper's own admissions, and they land on the central claim.\n\nThe soft spots are in proportion: the reconstruction part is solid; the generation evaluation is not. Also missing: no code/data yet, no baseline comparison to MatterGen or WyckoffTransformer, and the dataset filters (P1 excluded, ≤9 Wyckoff sites, etc.) limit coverage. The authors are honest about limitations, which helps, but honesty does not fill the gap.\n\nWho is this for? Someone working on crystal representation learning or generative models for materials will get value from the representation and reconstruction analysis. The generation claims need to be treated skeptically until code and validity-filter details are released.\n\nRecommendation: send to peer review rather than desk reject. The representation question is worth referee time. Ask the authors to release code, compare against baselines, and specify exactly how the validity filter handles partial occupancy. If that clarification fails, the validity numbers should be withdrawn.","headline":"Novel Wyckoff partial-occupancy representation is learnable, but the generation claim rests on an unspecified validity filter and weak property evidence.","tokens_in":17623,"tokens_out":2212,"would_cite":true,"duration_ms":22602,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Dis-GEN is the first generative model for disordered inorganic crystals, representing partial occupancies and vacancies directly at Wyckoff sites.","keywords":["disordered crystals","partial occupancy","Wyckoff sites","crystal structure generation","variational autoencoder","compositional disorder","vacancies","symmetry equivariance"],"falsifier":"Generate, say, one thousand structures that pass the symmetry and no-overlap filters, compute their energy above the convex hull at the generated compositions with a converged electronic-structure method, and compare with the paper's Zn-V-O result; if the overwhelming majority are thermodynamically unstable, the validity metrics used here do not establish physical plausibility.","tokens_in":16573,"feed_emoji":"💎","tokens_out":10061,"duration_ms":100652,"temperature":0.7,"pith_summary":"This paper introduces Dis-GEN, a variational autoencoder that learns disordered inorganic crystal structures from experimental entries in the largest crystal-structure database, and claims it is the first generative model able to produce such materials. Disordered crystals, where two or more atomic species share a crystallographic site or a site is partly vacant, are common in doped oxides, battery electrodes, and thermoelectrics, yet existing generative models assume every site is fully occupied. Dis-GEN represents each crystal by its space group, lattice parameters, and a matrix of Wyckoff sites whose rows encode partial occupancy, multiplicity, disorder status, fractional coordinates, and site letter, so both compositional disorder and vacancies can be sampled from a single average unit cell. The model reconstructs lattice parameters to about 0.06 Å and reaches 98.4 percent symmetry-consistency on the test set, with generated structures passing a 96 percent no-overlap validity filter; the authors also state that some generated compositions are chemically unreasonable.","feed_headline":"Dis-GEN generates disordered crystals without supercell enumeration","feed_subtitle":"A Wyckoff-site VAE trained on experimental crystal data captures partial occupancies and vacancies while preserving symmetry.","key_machinery":"The load-bearing object is the Wyckoff-site representation: each column of the atomic matrix corresponds to one Wyckoff site, a group of symmetry-equivalent positions in the unit cell, and stores the partial-occupancy distribution, multiplicity, disorder flag, fractional coordinates, and Wyckoff letter. Because space groups fix which Wyckoff letters and multiplicities are possible, the model can check its own reconstructions for internal symmetry consistency, and the paper uses this symmetry-matching accuracy as a filter that removes symmetry-violating structures before validity screening. The authors describe the representation as empirically equivariant: global rotations or translations that would change the encoding are reflected in the reconstruction behavior, while physical predictions are invariant. The VAE encoder processes the atomic matrix with a convolutional network and the crystal vector with a linear network, and the decoder emits seven outputs with separate loss terms; the partial-occupancy loss is a squared error between occupancy distributions, which is what lets the model learn site disorder and vacancies.","core_discovery":"The central claim is that disordered crystals can be generated without enumerating supercells. Dis-GEN encodes each structure as a crystal vector holding six lattice parameters and a one-hot space group, together with an atomic matrix whose columns are Wyckoff sites; each column carries one-hot encoded partial occupancies, Wyckoff multiplicity, a disorder indicator, fractional coordinates, and a Wyckoff letter. A variational autoencoder with convolutional and linear layers compresses this representation and reconstructs it through seven task-specific losses, so partial occupancy is treated as a distribution over species rather than a single label. On a held-out test set the model reconstructs lattice parameters to within a few hundredths of an angstrom, recovers the space group with 99.4 percent accuracy, and achieves 98.4 percent symmetry-matching accuracy, meaning the predicted Wyckoff letters and multiplicities agree with the predicted space group. Sampling from a kernel-density estimate of the latent space yields generated structures with roughly 99 percent symmetry-matching accuracy and 96 percent validity; the authors show example generations and a composition-conditioned search in the Zn-V-O system, while noting that the generated phases in that system are thermodynamically unstable.","pith_inferences":["A straightforward stress test would be to train the same VAE on a supercell-enumerated encoding of the same experimental entries; if the average-cell model matches the supercell model in downstream property averages, the independent-partial-occupancy assumption is validated.","Because only about half of generated structures pass a charge-neutrality filter, coupling the generator with an electronegativity- or charge-based chemical filter is likely to be needed before the outputs are useful for discovery; the paper notes such filters do not yet exist.","The representation's success at capturing partial occupancies suggests vacancies could be treated as a fourth 'species' with an occupancy value, but the paper does not evaluate vacancy energetics, so that extension remains untested."],"forward_implications":["If Dis-GEN works as claimed, a disordered material can be sampled as one average unit cell rather than as hundreds of supercells, changing the cost of exploring doped and vacancy-containing phases.","The Wyckoff-site representation is not tied to the variational autoencoder, so the same encoding could be used by diffusion or transformer generators to handle partial occupancies.","Composition-conditioned sampling from the latent space gives a concrete way to propose dopant configurations in chemistries that are sparse in experimental databases, even before properties are computed.","The reported reconstruction errors suggest the model produces CIF-like structures that are suitable inputs for downstream relaxation or property screening pipelines."],"supporting_citations":[{"why":"Supplies the structural validity criterion used to screen generated structures and the full-occupancy VAE baseline that Dis-GEN extends.","marker":"[3]"},{"why":"Prior generative model for disordered structures, limited to two atoms swapping at a site; Dis-GEN's representation of partial occupancy is positioned against it.","marker":"[6]"},{"why":"Establishes the Wyckoff-site VAE representation for fully ordered crystals that this work adapts to disordered sites.","marker":"[26]"},{"why":"Describes the experimental structure database that provides the training set.","marker":"[27]"},{"why":"Provides the DFT-derived structures and phase-diagram data used for comparison in the Zn-V-O case study.","marker":"[30]"},{"why":"Provides the symmetry-refinement procedure used to move generated fractional coordinates onto Wyckoff positions.","marker":"[35]"},{"why":"Supplies the coordinate-refinement procedure that Dis-GEN's reconstruction post-processing is modelled on.","marker":"[36]"},{"why":"Provides the machine-learned force field used to relax generated candidates before stability analysis.","marker":"[48]"},{"why":"Supplies the pretraining data for the force-field model used in the stability evaluation.","marker":"[49]"}],"fun_headline_variants":["Dis-GEN: disorder without supercell enumeration","Symmetry-consistent generation of disordered crystals","AI generates disordered crystals without supercells","Disordered crystal generation with Wyckoff sites","Dis-GEN: symmetry-preserving generative model for disorder"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a disordered crystal is adequately described by an average unit cell in which each Wyckoff site carries independent partial occupancies, and that a generated structure counts as physically plausible if it is symmetry-consistent and has no pair of atoms closer than 0.5 Å.","fun_headline_variants_meta":{"raw":{"variants":["Dis-GEN: disorder without supercell enumeration","Symmetry-consistent generation of disordered crystals","AI generates disordered crystals without supercells","Disordered crystal generation with Wyckoff sites","Dis-GEN: symmetry-preserving generative model for disorder"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000874,"raw_usage":{"total_tokens":3786,"prompt_tokens":955,"completion_tokens":2831,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":571,"completion_tokens_details":{"reasoning_tokens":2761}},"tokens_in":571,"tokens_out":2831,"duration_ms":22463,"temperature":1.0,"reasoning_tokens":2761,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:15:23.512672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate, say, one thousand structures that pass the symmetry and no-overlap filters, compute their energy above the convex hull at the generated compositions with a converged electronic-structure method, and compare with the paper's Zn-V-O result; if the overwhelming majority are thermodynamically unstable, the validity metrics used here do not establish physical plausibility.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Prior generative model for disordered structures, limited to two atoms swapping at a site; Dis-GEN's representation of partial occupancy is positioned against it."}],"review_version":1}